IOADA: An Optimal Automated Augmentation Algorithm for Medical Image Segmentation

Deep learning has seen remarkable advancements in the field of automatic data augmentation, but its implementation in medical imaging remains constrained due to a predominant focus on natural images. To solve the problem of complex spatial information and high acquisition cost of medical images, we propose a differentiable automatic data augmentation framework based on Sinkhorn optimal transport and proximal updates(IOADA), which encompasses a specialized search space and leverages enhanced algorithms tailored specifically for complex medical images. Through our approach, the Dice Similarity Coefficient (DSC) for retinal image segmentation increased by 2.18% and 3.78
^%
on two datasets. Additionally, the DSC scores on the SKIN, STS2D, and LIVER datasets all exceeded 90%, and the model significantly reduces the time costs involved. Furthermore, our approach demonstrates successful transferability to other medical image segmentation tasks. These findings underscore the potential of our method in enhancing medical image analysis by employing efficient and effective data augmentation techniques.